Digital watermarking ABSTRACT 1 GENERAL ... - CiteSeerX

25 downloads 2871 Views 442KB Size Report
This paper presents a process able to mark digital pictures with an invisible and ... allows the detection of watermarked pictures in a stream of digital images, only with ..... tication: Digital signature' from 'contemporary cryptology: the science of ...
Digital watermarking J.-F. Delaigle, C. De Vleeschouwer, B. Macq Laboratoire de Telecommunications et Teledetection Universite catholique de Louvain B^atiment Stevin - 2, place du Levant B-1348 Louvain-la-Neuve Tel.: +32 10 47.80.72 - Fax: +32 10 47.20.89 E-mail: [email protected]

ABSTRACT This paper presents a process able to mark digital pictures with an invisible and undetectable secrete information, called the watermark. This process can be the basis of a complete copyright protection system. The process rst step consists in producing a secrete image . The rst part of the secret resides in a basic information that forms a binary image. That picture is then frequency modulated. The second part of the secret is precisely the frequencies of the carriers. Both secrets depends on the identity of the copyright owner and on the original picture contents. The obtained picture is called the stamp. The second step consists in modulating the ampitude of the stamp according to a masking criterion stemming from a model of human perception. That too theoretical criterion is corrected by means of morphological tools helping to locate in the picture the places where the criterion is supposed not to match. This is followed by the adaptation of the level of the stamp at that places. The so formed watermark is then added to the original to ensure its protection. That watermarking method allows the detection of watermarked pictures in a stream of digital images, only with the knowledge of the picture owner's secrets.

Keywords: copyright protection, watermark, secrete key, masking, human vision model, perceptive components, morphology, robustness, detection, correlation.

1 GENERAL INTRODUCTION With the increasing availabilityof digitally stored information and the development of new multimedia services, security questions are becoming even more urgent. The acceptance of new services depends on whether suitable techniques for the protection of the work providers' interests are available.1 Moreover the nature of digital media threatens its own viability:

 First the replication of digital works is very easy and, what is more dangerous, really perfect. The copy is identical to the original.

 The ease of transmission and multiple uses is very worrying, too. Once a single pirate copy has been made, it is instantaneously accessible to anyone who wants it, without any control of the original picture owner.  Eventually the plasticity of digital media is a great menace. Any malevolent user (a pirate) can modify an image at will. Such maniplations are really easy for a pirate and put many copyright protection methods at risk.

According to these considerations the conception of a copyright protection system is really vital and it constitutes a great challenge, because it should cope with all these threats. Without watermarking, most authors will not dare to broadcast their work. This paper presents an additive watermarking technique. It consits in producing a synthetic picture (also called the stamp) which holds informations about the ownership of the original image and depends on the picture contents. That stamp is added to the original in a way that resulting picture is perceptually identical to the original one and so that the stamp is indetectable by a pirate computer. The aim of that technique is not the authentication of the picture content nor the identication of the owner. It is to allow a controller (i.e. the owner's computer or a Trusted Third Part) to nd out watermarked pictures in a stream of images with the knowledge of the owner's secret key in order to detect broadcast of illegal copies. The most interesting part of that method is the embedding process i.e. the weighting of each pixels of the stamp before adding it to the original. This is based on the masking concept coming from a model of human vision (the perceptive model). From this concept was deduced a method which reveals itself actually ecient. Another interesting part is the presentation of two methods used for the detection of watermarked pictures without the original. This last point is fundemental for the management of the copyright protection. Eventually this paper ends with the analyse of the results and the system robustness.

2 THE MASKING 2.1 Introduction The aim of a watermarking technique is to provide an invisible embedding of a secrete information, the watermark. This watermark must be masked (hidden) by the picture it is inlayed in. Precisely a master thesis has lead to a masking criterion deduced from physiological and psychophysic studies.2 Nevertheless, this theoretical criterion having been formulated for monochromatic signals, it had to be adaptated to suit real images.

2.2 The perceptive model: approximation of the eye functionment It is now admitted that the retina of the eye splits an image in several components. These components circulate from the eye to the cortex by di erent tuned channels, one channel being tuned to one component. The characteristics of one component are:

 the location in the visual eld (in the image).  the spatial frequency (in the Fourier domain: the amplitude in polar coordinates).  the orientation (in the Fourier domain: the phase in polar coordinates)

So, one perceptive channel can only be excited by one component of a signal whose characteristics are tuned to its. Components that have di erent characteristics are independent.

2.3 The masking concept According to perceptive model of human vision,3 signals that have same (near) components take the same channels from the eye to the cortex. It appears that such signals interact and are submitted to non-linear e ects. The masking is one of those e ects.

De nition: the detection threshold is the minimum level below which a signal can not be seen. De nition: the masking occurs when the detection threshold is increased because of the presence of another

signal.

In other words, there is masking when a signal can not be seen because of another with near characteristics and at a higher level.

2.4 The masking model With the object of modalizing the masking phenomenon, tests have been made on monochromatic signals, also called gratings. It appears that the eye is sensitive to the contrast of those gratings. This contrast is de ned by: ; Lmin) C = 2(Lmax (1) Lmax + Lmin where L is the luminance. It is possible to determine experimentally the detection threshold of one signal of contrast Cs with respect to the contrast Cm of the masking signal. That threshold can be modalized as follows: logCs

6

; ; ; ; ;  ; ; ; ;; logCm

-

Such bilogarithmic curves are traced for signals of one single frequency and one orientation (f0 , 0 ). The expression of the detection threshold is thus: Cs = max[C0; C0( CCm ) ] 0 where  (the slope) depends on (f0 , 0 ), typically, 0:6    1:1.

(2)

It is possible to extend that expression to introduce frequency dependence. The general expression of the detection threshold is becomes:

where:

Cs(Cm ; f; ) = C0 + k(f0 ;0 ) (f; )[Cs (f0 ;0 ) (Cm ) ; C0 ]

(3)

2 log2 ( f ) k(f0 ;0 ) (f; ) = exp[;( F 2(ff0) + (;2 (f0 )) ] 0 0

(4)

In that expression, f0 and 0 are relevant to the masking signal, f and  are relevant to the masked signal, F(f0 ) and (f0 ) are parameters that represent the spreading of the Gaussian function, C0 is often negligable. The spread of the gaussian function depends upon the frequency f0 : For frequency, typical bandwith at half response are 2,5 octaves at 1 c/d and 1,5 octaves at 16 c/d with a linear decrease between both frequencies.4 For orientation, half bandwith at half response depends on f0 and it takes typical values like 30 degrees at 1 c/d and 15 degrees at 16 c/d.5 After this expression, the frequency dependence of the detection threshold has a Gaussian form. Only near frequency signals can interact. When the frequency of the masking signal (the mask) is far from this of the signal to mask, the detection threshold is almost equal to C0.

2.5 The masking criterion It is important to notice that those results concern only gratings signals. To deduce a masking criterion that will apply to signals like real images, the preceding masking condition has to be adaptated. So, it is necessary to de ne a new concept able to take the place of the contrast, because the contrast is not de ne for real images. That new concept,2 is the local energy. The local energy is de ned on narrowband signals centered around one frequency and one orientation. A picture which is a broadband signal is rst ltered by Gabor narrowband lters, whose characteristics are near to human perception. The local energy around one frequency and one orientation is calculated following the scheme presented in this gure: I(x ,y)

- analytic lters (f ;  ) 0

0

- j:j

2

- local energy f ; ( 0

0)

(x, y)=E(f0 ;0 ) (x; y)

The masking criterion: If the local energy of one picture is less than the local energy of the mask, around all the frequencies (f0 ; 0) and for each pixel (x, y), then one can say that the picture is masked by the mask. Strictly, a picture is masked by a mask if 8(x; y) and 8(f0 ; 0 ); Emask;(f0;0 ) (x; y)  Epicture;(f0;0 ) (x; y). For real images, a good approximation of this criterion can be obtained by using a bank of lters whose central frequencies correspond to independent components and which are spread on all the Fourier space. It is admitted that 4 or 5 frequencies and 4 to 9 orientations are sucient. The standard choice is twenty lters (5 frequencies and 4 orientations).

Figure 1: Example of basic information used

2.6 Conclusion This section has lead to the expression of an easily implementable masking criterion appliable to any image. But this criterion is only an extension of a theoretic criterion appliable to monochromatic signals. Thus cases where that criterion does not match are possible.

3 PRINCIPLE OF THE SYSTEM 3.1 Basic information of the watermark This information is a binary picture looking like a modi ed checkerboard ( gure 1). As explained later, the pixels value of the square forming that picture can correspond to a binary sequence deduced from the copyright owner's (CO) secrete key.

3.2 The stamp In order to take advantage of the eye behaviour, the basic information is modulated at di erent frequencies and orientations corresponding to rather independent components. Moreover, we take care to lter the initial checkerboard with a low pass lter (LPF) (i.e. a Butterworth LPF) so that the resulting signal is bandlimited. This point is very important because it permits to limit the veri cation of the masking criterion in the coresponding channel. The position of the modulating carriers is secret. It can be deduced from CO's secret key. In practice, the frequency plan is divided into sectors. Each sector is relevant to one perceptive component and de ned a group of couples (f; ) where basic information can be modulated. Only one couple is chosen for each sector (because couples of a same sector don't stimulate independent components). The picture obtained from the sum of each modulated grid is called the stamp S(x; y). S(x; y) =

X G(x; y):cos(f

j 2K

xj :x + fyj :y)

(5)

K represents the set of sectors and (fx ; fy ) correspond to the couple chosen in sector j ( this couple is designed by the CO's secrete key). j

j

3.3 The position of the process in a global copyright scheme The process should be placed in a copyright protection scheme like drawn at gure 2. The skeletization function consists in an image processing program extracting essential characteristics from an image. The result is a bitsteam. This must be followed by a hash-function6 whose result is a succession of blocks of bits. Every block has the same length. The skeletization function gives the same result for two near images (i.e. original image and watermarked image). But the H-function always gives di erent results from di erent bitstreams as inputs. So, the inscription keys will be di erent for perceptually distinct pictures. After the Hfunction, the ciphering function is a trapdoor function.6 Thanks to this function the inscription keys used to deduce the basic grid and the position of the carriers depends on the CO's secret key. The aim of the use of a trapdoor function is to prevent someone from reproducing the same inscription keys with the knowledge of the H-function result. But it is possible for anyone to inverse that trapdoor function and to nd the H-function result from the inscription keys. It can be interesting in a proof procedure.

4 IMPLEMENTATION 4.1 Inscription The purpose of the inscription is to adapt the level of each part of the stamp ( for all frequencies ) to make it invisible once added to the picture. As mentioned above, each part of the stamp is narrow band. Inscriptions at di erent frequencies are thus independent and one can treat the di erent components of the stamp one at a time. For each frequency designed by the inscription keys, the procedure is divided in three steps : the modulation, the regulation of the level and the correction.

 Modulation

The rst step consists in the modulation of the particular carrier by the lowpass grid G(x; y). The result is G(x; y):cos(fx :x + fy :y), where fx and fy are the carrier position.  Regulation of the level According to the perceptual model, in order to guarantee the invisibility of the watermark its local energy has to be inferior to the picture local energy for each pixel around the inscription frequency. A way to reach this objective is to multiplicate the modulated grid by a weighting mask Weightj (x; y) reducing the amplitude of the stamp where energy in the correponding component of the original picture is weak. Nevertheless, one must take care to keep the narrow band characteristic of the resulting signal Sj (x; y) (= Weightj (x; y):G(x; y):cos(fx :x+fy :y)) in order to avoid non linear interactions between di erent parts of the stamp. In conclusion, 8j, we have to nd a signal Weightj (x; y) so that: { 8(x; y) ES (x; y) < EI;(f ;f )(x; y) { Sj is narrow band For simpli cation, lets consider W eightj (x; y) be composed of two factors: { j , a constant factor ( xing the global level of the stamp). { Mj (x; y), a mask whose values 2 [0; 1]. When j is chosen, the way to nd Mj (x; y) so that Weightj (x; y) satisfy the conditions de ned above is the following: j

j

j

j

j

j

xj

xj

j

Author’s secret key Picture

Skeletization

Ciphering

function

function

Inscription key (f,θ) where we write

What we write:

INSCRIPTION (with masking)

C H A N N E L

Author’s secret key Eventually:

Skeletoning

Ciphering

Segmentation

function

function

tool Inscription key

RECEPTION

Filtering + demodulation C = ∑ [G (x y ) I R (x y )] Correlation ,

(x ,y )

YES

NO

Figure 2: Global scheme for copyright protection.

.

,

{ Firstly, Mj (x; y) is a binary mask. Mj (x; y) = 1 when the local energy of the stamp pemits the masking

and Mj (x; y) = 0 when the local energy of the stamp is too important. It is obvious that the initial choice of j has a direct in uence on Mj (x; y). Indeed, a great j value will lead to put most of the Mj (x; y) values to zero, while a small j value will lead to keep most of Mj (x; y) values at one. { Secondly, Weightj (x; y) is ltered so that the stamp remains narrow band. { After this second step, one has found a signal j :Mj (x; y):G(x; y) which is better masked than j :G(x; y). In order to really satisfy the masking criterion 8(x; y), this procedure must be repeated iteratively, taking Mj (x; y):G(x; y) as new G(x; y). Experiments have shown that only two iterations are sucient to have a result satisfying the masking criterion everywhere.

One important question remains: how to choose j ?

It has already been said that the more j increases, the more Mj (x; y) has points equal to zero. A trade o has P to be found by means of a de ned criterion. Maximizing the correlation at the detection (by maximizing j :Mj (x; y):G(x; y)) could have been a good criterion, but such a criterion often tends to impose an optimum with a lot of points equal to zero and a small number of points with a great value. The addition of the so obtained watermark generally entails a degradation of the picture quality. This emphasizes the lack of the masking criterion used. As mentioned in section 2.6, the invisibility criterion used here is an extension for real images. It appears that this extension entails some imperfections. This criterion being insucient, some improvements have been brought thanks to experimental results. The conclusion of these observations is that the invisibility is only strictly observed in high activity regions, where the local energy of high fequencies is important. These regions have to be favoured during the inscription in the sense that the level of the watermark will be increased in those regions while it has to be decreased in other regions. The correction process rst isolates the high activity regions ( gure 3.a). Then, an homogeneization of this picture is performed by use of morphological tools, e.g. one opening and one closing ( gure 3.b). After a leveling (in fact, a division by the mean or mean square value of the homogenized mask), we obtain a new mask used to multiply the picture local energy and so, giving an advantage to regions of highfrequency energy in comparison with other areas. After that correction, the process is identical to the one described previously. Moreover, the complexity is not increased. Indeed, we rst work on the inscription at high frequencies (where there is no quality problems). The value of high frequency local energy is then used for the calculation of the correcting mask used for inscription at lower frequencies. The correction scheme is drawn in the following schema.  HF energy

- opening

-

closing

- leveling -shaping MASK

4.2 Detection The aim is to detect if a watermark has been embedded. This can be done with the use of a correlation, but rst it is necessary to isolate the watermark and then to demodulate it in order to reconstruct something that is highly correlated with the basic information (the grid). The formulation of the watermark is: W(x; y) =

X A :cos(f

j 2K

j

xj :x + fyj :y)

(6)

(b)

(a)

Figure 3: Correcting mask for Lena: (a) Areas of high frequencies, (b) Morphological homogeneization of the mask. whereAj = j :G(x; y):M(x; y) (7) In this expression, M(x; y) adjusts the level of the grid in order it becomes invisible, it is called a mask, and its maximal value is one. j is a constant that used to normalize the mask, it must be as high as possible. The detection is divided in three steps : teh demodulation, the correlation and the decision.

 Demodulation

IW (x; y) =

X A :cos(f j

j 2K

xj :x + fyj :y) + IO + N(x; y)

(8)

where IW (x; y) is the watermarked picture, IO (x; y) is the original picture and N(x; y) is an additive noise from the channel. The demodulation consists in multiplying IW by cos(fx :x + fy :y); 8j 2 K and then to lter with a LP lter. The result will be : (9) Dj (x; y) = 12 :Aj (x; y) + N ? (x; y) N ? (x; y) depends on the image and on the additive noise. The other parts of the stamp will be eliminated by the LP lter. P  Correlation It consists in mutiplying the demodulated information D(x; y) = j 2K Dj (x; y) with the basic grid G(x; y). If the picture has not been too deteriorated, D(x; y) and G(x; y) should be similar. j

C =

X X D (x; y):G(x; y)

j 2K x;y

j

j

(10)

=

X X[G (x; y):M (x; y) + G(x; y):N ?(x; y)]

j 2K

j

2

x;y

j

In 11, the rst term is even greater than the second, because G and N ? have null average values. So C exclusively depends on the watermark value. in the case the grid is not the good one, the correlation gives: X X C ? = j G(x; y):G?(x; y)::Mj (x; y) j 2K

x;y

(11)

(12)

C ?  C if the choice of the basic information has been appropriate.  decision The detection algorithm performs demodulations and correlations at diverse frequencies and with diverse grids.The decision is made after the comparison of these correlations.

5 RESULTS The rst and probably mosty important result is the invisibility of the stam in all images that were tested. Figure 4.a and b compares the original and stamped picture for Lena. In gure 4.e, omne observes the watermark that was added to the original picture. Two methods were used to determine whether an image is watermarked or not. The rst one consists in comparing the result of C the correlation made with the right grid G(x; y) from the right key with C ? the correlation made with G? (x; y), the grid obtained by random keys see 12. If the picture is watermarked, the correlation with the right key is even greater than the random correlations. The results below (Figure 5) show the pertinence of this method. The second method uses a grid G(x; y) formed from a MLS sequence, having good correlation properties. Correlations are made with shifted versions of the basic grid. Due to these good correlation properties, the correlation with the the right grid gives a result even greater than the correlations with shifted grids. Results are presented below ( gure 4.c and d), if a picture is watermarked, a pick appears in the center.

6 SYSTEM ROBUSTNESS Many tests have been performed concerning usual pictures deteriorations in image processing like blurring and compression. The inspection of these results are quite satisfying, but expected due to the frequency approach. For all classical pirate attacks like zoom, cropping, overwatermarking it is not as simple. The overwatermarking makes no problem, the presence of the watermark is still detected. But for zoom and cropping, the remaing point is to nd a few tools permitting to complete the process. The concept of these tools is already de ned but yet no implementation has been acheived.7

7 CONCLUSION The process developed here allows the watermarking of the ownership of any picture. The perceptual approach used here is probably the best one, that is why the results obtained are so satifying compared with other methods and this method is so performant. Nevertheless studies are still running to acheive a new goal, consisting in

(a)

(b)

(c)

(d)

(e)

Figure 4: Results for Lena: (a) Original, (b) Watermarked one, (c) Correlation gra c for original, (d) Correlation gra c for watermarked, (e) Watermark.

Image Name

Optimal

Random

Random

Random

Random

correlation

correlation 1

correlation 2

correlation3

correlation 4

584609

92605

133920

80534

143633

94538

98099

135492

76739

137120

Conclusion

Lena watermarked Lena original

watermarked Non watermarked

Figure 5: Results of correlation for Lena and decision. making more information (e.g. ownership, date of marking) readable by the key owner from the watermark. This could be useful for real copyright protection protocols8 .9

8 REFERENCES [1] Kahin B. The strategic environment for protecting multimedia. volume 1, pages 1{8. IMA Intellectual Property Project Proceedings, January 1994. [2] Comes S. Les traitements perceptifs d'images numerisees. PhD thesis, Universite Catholique de Louvain, June 1995. [3] Olzak L.A. and Thomas J.P. Handbook of perception and human performance vol.1: Seeing spatial patterns. chapter 7. [4] G.C. Phillips H.R. Wilson, D.K. McFarlane. Spatial frequency tuning of orientation selective units estimated by oblique masking. Vision Research, 23(9):873{847, 1983. [5] G.C. Phillips H.R. Wilson. Orientation bandwiths of spatial mechanisms measured by masking. J. Opt. Soc. Am. A, 1(2):226{232, February 1984. [6] Edited by Gustavus J. Simmons. Section 1: Chapter 4: 'public key cryptography' and section 2: Chapter 6: 'authentication: Digital signature' from 'contemporary cryptology: the science of information integrity' ieee press, 1992. [7] J.F. Delaigle and C. De Vleeschouwer. Etiquetage d'images numeriques en vue de la protection des droits d'auteur, Juin 1995. [8] J.F. Delaigle C. Simon and B. Macq. Talisman (ac019): Technical state of the art. January 1996. [9] O. Bruyndonckx J.M. Boucqueau and B. Macq. Watermarking: workpackage 5 of accopi. June 1995.